OpenAI Codex Adds Security Scans and Ultrafast Tier at DevDay

What OpenAI shipped

At DevDay 2026 in San Francisco, OpenAI made Codex a larger product. Codex now has reusable cloud environments, a refreshed CLI with voice control, a code review view in the ChatGPT desktop app, and an automated vulnerability scanner called Codex Security Cloud.21 The company also announced changes to its API. The Agents API can now use Computer Use, a new Decisions API handles quick single-answer choices, and a premium speed tier called Ultrafast costs six times the standard API rate.21

The coding-assistant market has been shipping updates almost weekly, so the scope of the release matters. Each feature lines up with a recent move by Anthropic, GitHub, or Cursor. Read together, these updates show the competition has moved past which model writes better code. The fight now is over who controls review, security, and orchestration for code that agents produce.

Codex Security Cloud

The most consequential launch is the security scanner. Codex Security Cloud scans whole GitHub repositories on demand or on a schedule and can watch new commits as they arrive. Codex investigates each finding, removes duplicates, and prepares fixes in the cloud, even when the user's laptop is closed.21 It builds on the open-source Codex Security CLI, which OpenAI released in July for command-line and CI/CD use.21

One detailed walkthrough lists four stages. The tool builds a threat model, scans the code, tries to reproduce likely vulnerabilities in a sandbox to rule out false positives, and then proposes a fix without applying it. A developer reviews the generated patch before it becomes a draft pull request.25 OpenAI presents the tool as a complement to static application security testing (SAST), not a replacement.25 At a DevDay session, presenters said an internal security sprint fixed 53 critical findings on its first day. They also said a verify-fix step, which acts as an adversarial agent against each patch, kept the rollback rate to about 1%.26

The tool also opens access. Cloud users get OpenAI's Daybreak Blue defensive security models without applying to the Daybreak program, which used to require identity verification.21 It is available on Pro, Business, Enterprise, and Edu plans, on desktop and web.2128

Coverage of how the scanner works is not consistent. One forum write-up described it as a check on AI-generated code before merge that catches injection flaws and unsafe data handling.22 The more detailed reports describe something else: a scanner that covers entire repositories and keeps running.2125 The repository-level description is better supported, and it is the more important one. Checking each diff is already common. Running continuous, deduplicated scans that end in a reviewable fix is a product security teams would pay for.

Decisions API and Ultrafast

The Decisions API puts OpenAI into the market for fast, narrow "System One" models. Developers write a question with a fixed set of possible answers, supply text or images as context, and get one answer back. Typical uses are classification, request routing, and choosing an agent's next step.21 It runs on a version of GPT-6 Luna, OpenAI's low-cost model. OpenAI says it answers in about 150 milliseconds, compared with 1.6 seconds through the regular Luna API.21

Several observers connected the launch to Jev, a decision model that TypeSafe AI released less than two weeks earlier.2621 One Hacker News commenter called the Decisions API a validation of the category Jev started.30 Some descriptions at launch were loose. One write-up said it helps agents "weigh trade-offs."22 The narrower framing of a fixed answer set is the accurate one.27 It went from limited preview at DevDay to beta with GPT-6 Luna in early October.29

Ultrafast has the most caveats. OpenAI says it speeds up Codex token generation by up to eight times, to 300 tokens per second, and API generation by up to six times.21 The API price is six times the standard rate.21 One recap noted that the multiplier came from Sam Altman's remarks on stage, as quoted in a live blog, and was not in OpenAI's written summary.2826 At launch, access was limited to GPT-6 Astra for Enterprise customers and the new $500-a-month Pro 500 plan.21 GPT-6.1 Sol Ultrafast reached all API users on October 8.29 In Codex, Ultrafast uses up included subscription allowances eight times faster than standard.29

The pricing changes around Ultrafast deserve more attention than they got. The $200 Pro plan now includes ten times Plus usage, down from twenty. Weekly GPT-6 Pro messages drop from 200 to 100, though existing subscribers keep their current limits for a transition period.21 The new Pro 500 plan offers 25 times Plus usage.24 The pattern is to charge more for speed and give the cheaper heavy-user plan less.

Code review and security are where the competition is

OpenAI is not alone here. Anthropic's Code Review, which reviews an organization's pull requests, became generally available on October 15 with configurable review levels.37 Claude Code also released a series of point versions in October. These added Claude Mods, which let plugins change deeper behavior, and an opt-in side agent that flags problems the user or Claude might miss.3140 Many of those releases were mostly fixes: one version closed thirteen sandbox and permission escapes at once.38 Anthropic and OpenAI shipped competing CLI updates within 48 hours of each other in early October, and the focus was operations, such as worktree isolation and model pinning, more than model quality.3934

GitHub answered at Universe on October 28. It launched Copilot code review Max, a deeper review tier that GitHub says scored the highest F1 for high-severity issues on ReviewBench.13 It made Copilot Fusion, its model routing and orchestration system, generally available.11 It also released a public-preview security extension that brings code scanning alerts, Copilot explanations and fixes, and reviews of local changes into the Copilot app.13 The New Stack reported that GitHub claims Fusion beats Claude Code at about half the API cost per task.18 That is a vendor benchmark and should be read that way. GitHub also changed how it sells Copilot: Enterprise customers can now use all Copilot features, metered by AI credits, without a separate subscription.11

Cursor got to automated review earlier. On September 23 it added two bots: Rollouts, which watches deployments, and Security Review.3 Its April Cursor 3 redesign was built around running several agents in parallel,2 and its Pro plan still costs $20 a month.2

How the model lineups overlap

The vendors depend on each other's models. GitHub retired several older OpenAI models on October 19 and listed GPT-6.1 Sol and GPT-6 Luna as replacements.12 One outlet reported the replacements as "GPT-5.6" models. That conflicts with GitHub's own changelog, which should be treated as the authority.1712 OpenAI's newest models now run inside a competitor's product. That gives OpenAI distribution, and it also explains why Codex is adding features that keep developers in OpenAI's own interface.

Assessment

The DevDay launch is best understood as OpenAI's attempt to own the review-and-remediation stage of the development cycle. That stage is becoming the bottleneck: agents can now write code faster than people can review it. Codex Security Cloud is a meaningful product. Reproducing findings in a sandbox and checking each fix adversarially address the false-positive problem that makes security teams wary of automated scanners.2526

The overall bundle has weaknesses. Ultrafast is expensive and was gated at launch, and the Decisions API looks like a quick response to a startup. Developers also lost some value: the $200 plan's allowance was cut.21 Some attendees were not impressed. One Hacker News commenter said a rival Anthropic model release hung over the event.30

Within a month, GitHub had matched many of the announcements with a deeper review tier, a security extension, and model routing, and Anthropic's Code Review had reached general availability.1337 Being first with a feature lasts only weeks in this market. OpenAI is more likely to keep an advantage through the quality of its security models and how it prices access to them than through any single feature.